Bindel, Fall 2016 Matrix Computations (CS 6210) Notes for

Transcription

1 1 Logistics Notes for General announcement: we are switching from weekly to bi-weekly homeworks (mostly because the course is much bigger than planned). If you want to do HW but are not formally enrolled, send me to add you to CMS. Piazza will be set up today (not yet). Most of mathematics is best learned by doing. Linear algebra is no exception. You have had a previous class in which you learned the basics of linear algebra, and you will have plenty of practice with these concepts over the semester. This brief refresher lecture is supposed to remind you of what you ve already learned and introduce a few things you may not have seen. It also serves to set notation that will be used throughout the class. In addition to these notes, you may find it useful to go back to a good linear algebra text (there are several listed on the course syllabus) or to look at the linear algebra review chapters in the book. 2 Vectors A vector space (or linear space) is a set of vectors that can be added or scaled in a sensible way that is, addition is associative and commutative and scaling is distributive. We will generally denote vector spaces by script letters (e.g. V, W), vectors by lower case Roman letters (e.g. v, w), and scalars by lower case Greek letters (e.g. α, β). But we feel free to violate these conventions according to the dictates of our conscience or in deference to other conflicting conventions. There are many types of vector spaces. Apart from the ubiquitous spaces R n and C n, the most common vector spaces in applied mathematics are

2 different types of function spaces. These include P d = {polynomials of degree at most d}; V = {linear functions V R (or C)}; L(V, W) = {linear maps V W}; C k (Ω) = { k-times differentiable functions on a set Ω}; and many more. We compute with vectors in R n and C n, which we represent concretely by tuples of numbers in memory, usually stored in sequence. To keep a broader perspective, though, we will also frequently describe examples involving the polynomial spaces P d. 2.1 Spanning sets and bases We often think of a matrix as a set of vectors V = [ v 1 v 2... v n ]. The range space R(V ) or the span sp{v j } n j=1 is the set of vectors {V c : c R n } (or C n for complex spaces). The vectors are linearly independent if any vector in the span has a unique representation as a linear combination of the spanning vectors; equivalently, the vectors are linearly independent if there is no linear combination of the spanning vectors that equals zero except the one in which all coefficients are zero. A linearly independent set of vectors is a basis for a space V if sp{v j } n j=1 = V; the number of basis vectors n is the dimension of the space. Spaces like C k (Ω) do not generally have a finite basis; they are infinite-dimensional. We will focus on finite-dimensional spaces in the class, but it is useful to know that there are interesting infinitedimensional spaces in the broader world. The standard basis in R n is the set of column vectors e j with a one in the jth position and zeros elsewhere: I = [ e 1 e 2... e n ]. Of course, this is not the only basis. Any other set of n linearly independent vectors in R n is also a basis V = [ v 1 v 2... v n ].

3 The matrix V formed in this way is invertible, with multiplication by V corresponding to a change from the {v j } basis into the standard basis and multiplication by V 1 corresponding to a map from the {v j } basis to the standard basis. We will use matrix-like notation to describe bases and spanning sets even when we deal with spaces other than R n. For example, a common basis for P d is the power basis X = [ 1 x x 2... x d]. Each column in X is really a function of one variable (x), and matrixvector multiplication with X represents a map from a coefficient vector in R d+1 to a polynomial in P d. That is, we write polynomials p P d in terms of the basis as d p = Xc = c j x j and, we think of computing the coefficients from the abstract polynomial via a formal inverse: c = X 1 p. We typically think of a map like Y = X 1 in terms of rows j=0 y 0 Y y 1 =. where each row yj is a linear functional or dual vector (i.e. linear mappings from the vector space to the real or complex numbers). Collectively, {y0,..., yd } are the dual basis to {1, x,..., xd }. The power basis is not the only basis for P d. Other common choices include Newton or Lagrange polynomials with respect to a set of points, which you may have seen in another class such as CS In this class, we will sometimes use the Chebyshev 1 polynomial basis {T j (x)} given by the 1 Pafnuty Chebyshev was a nineteenth century Russian mathematician, and his name has been transliterated from the Cyrillic alphabet into the Latin alphabet in several different ways. We inherit our usual spelling from one of the French transliterations, but the symbol T for the polynomials comes from the German transliteration Tschebyscheff. y d

4 recurrence T 0 (x) = 1 T 1 (x) = x T j+1 (x) = 2xT j (x) T j 1 (x), j 1, and the Legendre polynomial basis {P j (x)}, given by the recurrence P 0 (x) = 1 P 1 (x) = x (j + 1)P j+1 (x) = (2j + 1)xP j (x) jp j 1 (x). As we will see over the course of the semester, sometimes the obvious choice of basis (e.g. the standard basis in R n or the power basis in P d ) is not the best choice for numerical computations. 2.2 Vector norms A norm measures vector lengths. It is positive definite, homogeneous, and sub-additive: v 0 and v = 0 iff v = 0 αv = α v u + v u + v. The three most common vector norms we work with in R n are the Euclidean norm (aka the 2-norm), the -norm (or max norm), and the 1-norm: v 2 = v j 2 v = max v j j v 1 = j j v j Many other norms can be related to one of these three norms. In particular, a natural norm in an abstract vector space will often look strange in the corresponding concrete representation with respect to some basis function. For example, consider the vector space of polynomials with degree at

5 most 2 on [ 1, 1]. This space also has a natural Euclidean norm, max norm, and 1-norm; for a given polynomial p(x) these are 1 p 2 = p(x) 2 dx 1 p = max p 1 = p(x) x [ 1,1] 1 1 p(x) dx. But when we write p(x) in terms of the coefficient vector with respect to the power basis (for example), the max norm of the polynomial is not the same as the max norm of the coefficient vector. In fact, if we consider a polynomial p(x) = c 0 + c 1 x, then the max norm of the polynomial p is the same as the one-norm of the coefficient vector the proof of which is left as a useful exercise to the reader. In a finite-dimensional vector space, all norms are equivalent: that is, if and are two norms on the same finite-dimensional vector space, then there exist constants c and C such that for any v in the space, c v v C v. Of course, there is no guarantee that the constants are small! An isometry is a mapping that preserves vector norms. For R n, the only isometries for the 1-norm and the -norm are permutations. For Euclidean space, though, there is a much richer set of isometries, represented by the orthogonal matrices (matrices s.t. Q Q = I). 2.3 Inner products An inner product, is a function from two vectors into the real numbers (or complex numbers for an complex vector space). It is positive definite, linear in the first slot, and symmetric (or Hermitian in the case of complex vectors); that is: v, v 0 and v, v = 0 iff v = 0 αu, w = α u, w and u + v, w = u, w + v, w u, v = v, u,

6 where the overbar in the latter case corresponds to complex conjugation. A vector space with an inner product is sometimes called an inner product space or a Euclidean space. Every inner product defines a corresponding norm v = v, v The inner product and the associated norm satisfy the Cauchy-Schwarz inequality u, v u v. The standard inner product on C n is x y = y x = n ȳ j x j. But the standard inner product is not the only inner product, just as the standard Euclidean norm is not the only norm. Just as norms allow us to reason about size, inner products let us reason about angles. In particular, we define the cosine of the angle θ between nonzero vectors v and w as cos θ = j=1 v, w v w. Returning to our example of a vector space of polynomials, the standard L 2 ([ 1, 1]) inner product is p, q L 2 ([ 1,1]) = 1 1 p(x) q(x) dx. If we express p and q with respect to a basis (e.g. the power basis), we fine that we can represent this inner product via a symmetric positive definite matrix. For example, let p(x) = c 0 +c 1 x+c 2 x 2 and let q(x) = d 0 +d 1 x+d 2 x 2. Then where d 0 p, q L 2 ([ 1,1]) = d 1 d 2 a ij = 1 1 x i 1 x j 1 dx = a 00 a 01 a 02 a 10 a 11 a 12 a 20 a 21 a 22 c 0 c 1 c 2 { 2/(i + j 1), = d Ac = c, d A i + j even 0, otherwise

7 The symmetric positive definite matrix A is what is sometimes called the Gram matrix for the basis {1, x, x 2 }. We say two vectors u and v are orthogonal with respect to an inner product if u, v = 0. If u and v are orthogonal, we have the Pythagorean theorem: u + v 2 = u + v, u + v = u, u + v, u + u, v + v, v = u 2 + v 2 Two vectors u and v are orthonormal if they are orthogonal with respect to the inner product and have unit length in the associated norm. When we work in an inner product space, we often use an orthonormal basis, i.e. a basis in which all the vectors are orthonormal. For example, the normalized Legendre polynomials 2 2j + 1 P j(x) form orthonormal bases for the P d inner product with respect to the L 2 ([ 1, 1]) inner product. 3 Matrices and mappings A matrix represents a mapping between two vector spaces. That is, if L : V W is a linear map, then the associated matrix A with respect to bases V and W satisfies A = W 1 LV. The same linear mapping corresponds to different matrices depending on the choices of basis. But matrices can reresent several other types of mappings as well. Over the course of this class, we will see several interpretations of matrices: Linear maps. A map L : V W is linear if L(x + y) = Lx + Ly and L(αx) = αlx. The corresponding matrix is A = W 1 LV. Linear operators. A linear map from a space to itself (L : V V) is a linear operator. The corresponding (square) matrix is A = V 1 LV. Bilinear forms. A map a : V W R (or C for complex spaces) is bilinear if it is linear in both slots: a(αu + v, w) = αa(u, w) + a(v, w) and a(v, αu + w) = αa(v, u) + a(v, w). The corresponding matrix has elements A ij = a(v i, w j ); if v = V c and w = W d then a(v, w) = d T Ac. We call a bilinear form on V V symmetric if a(v, w) = a(w, v); in this case, the corresponding matrix A is also symmetric (A = A T ). A

8 symmetric form and the corresponding matrix are called positive semidefinite if a(v, v) 0 for all v. The form and matrix are positive definite if a(v, v) > 0 for any v 0. A skew-symmetric matrix (A = A T ) corresponds to a skew-symmetric or anti-symmetric bilinear form, i.e. a(v, w) = a(w, v). Sesquilinear forms. A map a : V W C (where V and W are complex vector spaces) is sesquilinear if it is linear in the first slot and the conjugate is linear in the second slot: a(αu + v, w) = αa(u, w)+a(v, w) and a(v, αu+w) = ᾱa(v, u)+a(v, w). The matrix has elements A ij = a(v i, w j ); if v = V c and w = W d then a(v, w) = d Ac. We call a sesquilinear form on V V Hermitian if a(v, w) = a(w, v); in this case, the corresponding matrix A is also Hermitian (A = A ). A Hermitian form and the corresponding matrix are called positive semidefinite if a(v, v) 0 for all v. The form and matrix are positive definite if a(v, v) > 0 for any v 0. A skew-hermitian matrix (A = A ) corresponds to a skew-hermitian or anti-hermitian bilinear form, i.e. a(v, w) = a(w, v). Quadratic forms. A quadratic form φ : V R (or C) is a homogeneous quadratic function on V, i.e. φ(αv) = α 2 φ(v) for which the map b(v, w) = φ(v + w) φ(v) φ(w) is bilinear. Any quadratic form on a finite-dimensional space can be represented as c Ac where c is the coefficient vector for some Hermitian matrix A. The formula for the elements of A given φ is left as an exercise. We care about linear maps and linear operators almost everywhere, and most students come out of a first linear algebra class with some notion that these are important. But apart from very standard examples (inner products and norms), many students have only a vague notion of what a bilinear form, sesquilinear form, or quadratic form might be. Bilinear forms and sesquilinear forms show up when we discuss large-scale solvers based on projection methods. Quadratic forms are important in optimization, physics (where they often represent energy), and statistics (e.g. for understanding variance and covariance).

9 3.1 Matrix norms The space of matrices forms a vector space; and, as with other vector spaces, it makes sense to talk about norms. In particular, we frequently want norms that are consistent with vector norms on the range and domain spaces; that is, for any w and v, we want w = Av = w A v. One obvious consistent norm is the Frobenius norm, A = i,j a 2 ij. Even more useful are induced norms (or operator norms) Av A = sup v 0 v = sup Av. v =1 The induced norms corresponding to the vector 1-norm and -norm are A 1 = max a ij (max column sum) j i A = max a ij (max row sum) i j The norm induced by the vector Euclidean norm (variously called the matrix 2-norm or the spectral norm) is more complicated. The Frobenius norm and the matrix 2-norm are both orthogonally invariant (or unitarily invariant in a complex vector space. That is, if Q is a square matrix with Q = Q 1 (an orthogonal or unitary matrix) of the appropriate dimensions QA F = A F, AQ F = A F, QA 2 = A 2, AQ 2 = A 2. This property will turn out to be frequently useful throughout the course.

10 3.2 Decompositions and canonical forms Matrix decompositions (also known as matrix factorizations) are central to numerical linear algebra. We will get to know six such factorizations well: P A = LU (a.k.a. Gaussian elimination). Here L is unit lower triangular (triangular with 1 along the main diagonal), U is upper triangular, and P is a permutation matrix. A = LL (a.k.a. Cholesky factorization). Here A is Hermitian and positive definite, and L is a lower triangular matrix. A = QR (a.k.a. QR decomposition). Here Q has orthonormal columns and R is upper triangular. If we think of the columns of A as a basis, QR decomposition corresponds to the Gram-Schmidt orthogonalization process you have likely seen in the past (though we rarely compute with Gram-Schmidt). A = UΣV (a.k.a. the singular value decomposition or SVD). Here U and V have orthonormal columns and Σ is diagonal with non-negative entries. A = QΛQ (a.k.a. symmetric eigendecomposition). Here A is Hermitian (symmetric in the real case), Q is orthogonal or unitary, and Λ is a diagonal matrix with real numbers on the diagonal. A = QT Q (a.k.a. Schur form). Here A is a square matrix, Q is orthogonal or unitary, and T is upper triangular (or nearly so). The last three of these decompositions correspond to canonical forms for abstract operators. That is, we can view these decompositions as finding bases in which the matrix representation of some operator or form is particularly simple. More particularly: SVD: For any linear mapping L : V W, there are orthonormal bases for the two spaces such that the corresponding matrix is diagonal Symmetric eigendecomposition: For any Hermitian sesquilinear map on an inner product space, there is an orthonormal basis for the space such that the matrix representation is diagonal.

11 Schur form: For any linear operator L : V V, there is an orthonormal basis for the space such that the matrix representation is upper triangular. Equivalently, if {u 1,..., u n } is the basis in question, then sp({u j } k j=1) is an invariant subspace for each 1 k n. The Schur form turns out to be better for numerical work than the Jordan canonical form that you should have seen in an earlier class. We will discuss this in more detail when we discuss eigenvalue problems. 3.3 The SVD and the 2-norm The singular value decomposition is useful for a variety of reasons; we close off the lecture by showing one such use. Suppose A = UΣV is the singular value decomposition of some matrix. Using orthogonal invariance (unitary invariance) of the 2-norm, we have i.e. A 2 = U AV 2 = Σ 2, A 2 = max v 2 =1 j σ j v j 2 vj 2. That is, the spectral norm is the largest weighted average of the singular values, which is the same as just the largest singular value. The small singular values also have a meaning. If A is a square, invertible matrix then A 1 2 = V Σ 1 U 2 = Σ 1 2, i.e. A 1 2 is the inverse of the smallest singular value of A. The smallest singular value of a nonsingular matrix A can also be interpreted as the distance to singularity : if σ n is the smallest singular value of A, then there is a matrix E such that E 2 = σ n and A + E is singular; and there is no such matrix with smaller norm. These facts about the singular value decomposition are worth pondering, as they will be particularly useful in the next lecture when we ponder sensitivity and conditioning.